Multinomial classification with class-conditional overlapping sparse feature groups

Regularized multinomial logistic model is widely used in multi-class classification problems. For high dimension data, various regularization methods achieving sparsity have been developed and applied successfully to many real-world applications such as bioinformatics, health informatics and text mining. In many cases there exist intrinsic group structures among the features. Incorporating the group information in the model can enhance model performance. In multi-class classification, different classes may relate to different feature groups. With these considerations, we propose a class-conditional regularization of the multinomial logistic model (CCSOGL) to enable the discovery of class-specific feature groups. To solve the model, we developed an efficient cyclic block coordinate descent based algorithm. We also apply our method to analyze real-world datasets to demonstrate its superior performance.